def __init__(self,
img_in_size=256,
world_size_in_img=256,
feature_channels=32,
ground_channels=3,
embed_size=40,
aux_ground=False,
freeze=False):
super(TopDownToEgoMap, self).__init__(img_in_size, world_size_in_img)
# Process images using a resnet to get a feature map
self.feature_net = ResNet13Light(feature_channels, down_pad=True)
self.aux_ground = aux_ground
if aux_ground:
self.lang_filter = MapLangSemanticFilter(embed_size,
feature_channels,
ground_channels)
enable_weight_saving(self.lang_filter,
"ground_filter",
alwaysfreeze=freeze)
enable_weight_saving(self.feature_net,
"feature_resnet_light",
alwaysfreeze=freeze)
def __init__(self, run_name=""):
super(ModelTrajectoryToAction, self).__init__()
self.model_name = "lsvd_action"
self.run_name = run_name
self.writer = LoggingSummaryWriter(log_dir="runs/" + run_name)
self.params = get_current_parameters()["ModelPVN"]
self.aux_weights = get_current_parameters()["AuxWeights"]
self.prof = SimpleProfiler(torch_sync=PROFILE, print=PROFILE)
self.iter = nn.Parameter(torch.zeros(1), requires_grad=False)
# Common
# --------------------------------------------------------------------------------------------------------------
self.map_transform_w_to_s = MapTransformerBase(
source_map_size=self.params["global_map_size"],
dest_map_size=self.params["local_map_size"],
world_size=self.params["world_size_px"])
self.map_transform_r_to_w = MapTransformerBase(
source_map_size=self.params["local_map_size"],
dest_map_size=self.params["global_map_size"],
world_size=self.params["world_size_px"])
# Output an action given the global semantic map
if self.params["map_to_action"] == "downsample2":
self.map_to_action = EgoMapToActionTriplet(
map_channels=self.params["map_to_act_channels"],
map_size=self.params["local_map_size"],
other_features_size=self.params["emb_size"])
elif self.params["map_to_action"] == "cropped":
self.map_to_action = CroppedMapToActionTriplet(
map_channels=self.params["map_to_act_channels"],
map_size=self.params["local_map_size"],
manual=self.params["manual_rule"],
path_only=self.params["action_in_path_only"],
recurrence=self.params["action_recurrence"])
self.spatialsoftmax = SpatialSoftmax2d()
self.gt_fill_missing = MapBatchFillMissing(
self.params["local_map_size"], self.params["world_size_px"])
# Don't freeze the trajectory to action weights, because it will be pre-trained during path-prediction training
# and finetuned on all timesteps end-to-end
enable_weight_saving(self.map_to_action,
"map_to_action",
alwaysfreeze=False,
neverfreeze=True)
self.action_loss = ActionLoss()
self.env_id = None
self.seg_idx = None
self.prev_instruction = None
self.seq_step = 0
self.get_act_start_pose = None
self.gt_labels = None
def __init__(self, name, channels_in=32, map_world_size=32, *inputs):
super(GoalAuxiliary2D, self).__init__(name, *inputs)
self.gather_2d = Gather2D()
self.channels_in = channels_in
self.map_world_size = map_world_size
self.goal_linear = nn.Linear(channels_in, 2)
enable_weight_saving(self.goal_linear, "aux_goal_linear_" + name)
self.loss = nn.CrossEntropyLoss()
self.accuracy_meter = MovingAverageMeter(10)
def __init__(self, name, feature_vec_len=32, num_classes=2, num_outputs=1, *inputs):
super(ClassAuxiliary, self).__init__(name, *inputs)
self.channels_in = feature_vec_len
self.num_classes = num_classes
self.num_outputs = num_outputs
self.cls_linear = nn.Linear(feature_vec_len, num_classes * num_outputs)
enable_weight_saving(self.cls_linear, "aux_class_linear_" + name)
self.loss = nn.CrossEntropyLoss()
self.meter_accuracy = MovingAverageMeter(10)
def __init__(self, name, feature_vec_len=32, num_classes=64, dropout=0, *inputs):
super(ClassAuxiliary2D, self).__init__(name, *inputs)
self.gather_2d = Gather2D()
self.channels_in = feature_vec_len
self.dropout = nn.Dropout(dropout)
self.num_classes = num_classes
self.cls_linear = nn.Linear(feature_vec_len, num_classes)
enable_weight_saving(self.cls_linear, "aux_class_linear_2d_" + name)
self.loss = nn.CrossEntropyLoss()
if self.name == "aux_grounding_map":
self.loss.weight = torch.tensor([0.5, 0.5])
self.meter_accuracy = MovingAverageMeter(10)
def __init__(self,
name,
world_size_px=32,
feature_vec_len=32,
num_classes=64,
dropout=0,
*inputs):
super(ClassAuxiliary2D, self).__init__(name, *inputs)
self.gather_2d = Gather2D()
self.channels_in = feature_vec_len
self.dropout = nn.Dropout(dropout)
self.num_classes = num_classes
self.world_size_px = world_size_px
self.cls_linear = nn.Linear(feature_vec_len, num_classes)
enable_weight_saving(self.cls_linear, "aux_class_linear_2d_" + name)
self.loss = nn.CrossEntropyLoss()
self.meter_accuracy = MovingAverageMeter(10)
def __init__(self, run_name="", model_class=MODEL_RSS,
aux_class_features=False, aux_grounding_features=False,
aux_class_map=False, aux_grounding_map=False, aux_goal_map=False,
aux_lang=False, aux_traj=False, rot_noise=False, pos_noise=False):
super(ModelTrajectoryTopDown, self).__init__()
self.model_name = "sm_trajectory" + str(model_class)
self.model_class = model_class
print("Init model of type: ", str(model_class))
self.run_name = run_name
self.writer = LoggingSummaryWriter(log_dir="runs/" + run_name)
self.params = get_current_parameters()["Model"]
self.aux_weights = get_current_parameters()["AuxWeights"]
self.prof = SimpleProfiler(torch_sync=PROFILE, print=PROFILE)
self.iter = nn.Parameter(torch.zeros(1), requires_grad=False)
# Auxiliary Objectives
self.use_aux_class_features = aux_class_features
self.use_aux_grounding_features = aux_grounding_features
self.use_aux_class_on_map = aux_class_map
self.use_aux_grounding_on_map = aux_grounding_map
self.use_aux_goal_on_map = aux_goal_map
self.use_aux_lang = aux_lang
self.use_aux_traj_on_map = aux_traj
self.use_aux_reg_map = self.aux_weights["regularize_map"]
self.use_rot_noise = rot_noise
self.use_pos_noise = pos_noise
# Path-pred FPV model definition
# --------------------------------------------------------------------------------------------------------------
self.img_to_features_w = FPVToGlobalMap(
source_map_size=self.params["global_map_size"], world_size_px=self.params["world_size_px"], world_size=self.params["world_size_m"],
res_channels=self.params["resnet_channels"], map_channels=self.params["feature_channels"],
img_w=self.params["img_w"], img_h=self.params["img_h"], img_dbg=IMG_DBG)
self.map_accumulator_w = LeakyIntegratorGlobalMap(source_map_size=self.params["global_map_size"], world_in_map_size=self.params["world_size_px"])
# Pre-process the accumulated map to do language grounding if necessary - in the world reference frame
if self.use_aux_grounding_on_map and not self.use_aux_grounding_features:
self.map_processor_a_w = LangFilterMapProcessor(
source_map_size=self.params["global_map_size"],
world_size=self.params["world_size_px"],
embed_size=self.params["emb_size"],
in_channels=self.params["feature_channels"],
out_channels=self.params["relevance_channels"],
spatial=False, cat_out=True)
else:
self.map_processor_a_w = IdentityMapProcessor(source_map_size=self.params["global_map_size"], world_size=self.params["world_size_px"])
if self.use_aux_goal_on_map:
self.map_processor_b_r = LangFilterMapProcessor(source_map_size=self.params["local_map_size"],
world_size=self.params["world_size_px"],
embed_size=self.params["emb_size"],
in_channels=self.params["relevance_channels"],
out_channels=self.params["goal_channels"],
spatial=True, cat_out=True)
else:
self.map_processor_b_r = IdentityMapProcessor(source_map_size=self.params["local_map_size"],
world_size=self.params["world_size_px"])
pred_channels = self.params["goal_channels"] + self.params["relevance_channels"]
# Common
# --------------------------------------------------------------------------------------------------------------
# Sentence Embedding
self.sentence_embedding = SentenceEmbeddingSimple(
self.params["word_emb_size"], self.params["emb_size"], self.params["emb_layers"])
self.map_transform_w_to_r = MapTransformerBase(source_map_size=self.params["global_map_size"],
dest_map_size=self.params["local_map_size"],
world_size=self.params["world_size_px"])
self.map_transform_r_to_w = MapTransformerBase(source_map_size=self.params["local_map_size"],
dest_map_size=self.params["global_map_size"],
world_size=self.params["world_size_px"])
# Batch select is used to drop and forget semantic maps at those timestaps that we're not planning in
self.batch_select = MapBatchSelect()
# Since we only have path predictions for some timesteps (the ones not dropped above), we use this to fill
# in the missing pieces by reorienting the past trajectory prediction into the frame of the current timestep
self.map_batch_fill_missing = MapBatchFillMissing(self.params["local_map_size"], self.params["world_size_px"])
# Passing true to freeze will freeze these weights regardless of whether they've been explicitly reloaded or not
enable_weight_saving(self.sentence_embedding, "sentence_embedding", alwaysfreeze=False)
# Output an action given the global semantic map
if self.params["map_to_action"] == "downsample2":
self.map_to_action = EgoMapToActionTriplet(
map_channels=self.params["map_to_act_channels"],
map_size=self.params["local_map_size"],
other_features_size=self.params["emb_size"])
elif self.params["map_to_action"] == "cropped":
self.map_to_action = CroppedMapToActionTriplet(
map_channels=self.params["map_to_act_channels"],
map_size=self.params["local_map_size"],
other_features_size=self.params["emb_size"]
)
# Don't freeze the trajectory to action weights, because it will be pre-trained during path-prediction training
# and finetuned on all timesteps end-to-end
enable_weight_saving(self.map_to_action, "map_to_action", alwaysfreeze=False, neverfreeze=True)
# Auxiliary Objectives
# --------------------------------------------------------------------------------------------------------------
# We add all auxiliaries that are necessary. The first argument is the auxiliary name, followed by parameters,
# followed by variable number of names of inputs. ModuleWithAuxiliaries will automatically collect these inputs
# that have been saved with keep_auxiliary_input() during execution
if aux_class_features:
self.add_auxiliary(ClassAuxiliary2D("aux_class", None, self.params["feature_channels"], self.params["num_landmarks"], self.params["dropout"],
"fpv_features", "lm_pos_fpv", "lm_indices"))
if aux_grounding_features:
self.add_auxiliary(ClassAuxiliary2D("aux_ground", None, self.params["relevance_channels"], 2, self.params["dropout"],
"fpv_features_g", "lm_pos_fpv", "lm_mentioned"))
if aux_class_map:
self.add_auxiliary(ClassAuxiliary2D("aux_class_map", self.params["world_size_px"], self.params["feature_channels"], self.params["num_landmarks"], self.params["dropout"],
"map_s_w_select", "lm_pos_map_select", "lm_indices_select"))
if aux_grounding_map:
self.add_auxiliary(ClassAuxiliary2D("aux_grounding_map", self.params["world_size_px"], self.params["relevance_channels"], 2, self.params["dropout"],
"map_a_w_select", "lm_pos_map_select", "lm_mentioned_select"))
if aux_goal_map:
self.add_auxiliary(GoalAuxiliary2D("aux_goal_map", self.params["goal_channels"], self.params["world_size_px"],
"map_b_w", "goal_pos_map"))
# RSS model uses templated data for landmark and side prediction
if self.use_aux_lang and self.params["templates"]:
self.add_auxiliary(ClassAuxiliary("aux_lang_lm", self.params["emb_size"], self.params["num_landmarks"], 1,
"sentence_embed", "lm_mentioned_tplt"))
self.add_auxiliary(ClassAuxiliary("aux_lang_side", self.params["emb_size"], self.params["num_sides"], 1,
"sentence_embed", "side_mentioned_tplt"))
# CoRL model uses alignment-model groundings
elif self.use_aux_lang:
# one output for each landmark, 2 classes per output. This is for finetuning, so use the embedding that's gonna be fine tuned
self.add_auxiliary(ClassAuxiliary("aux_lang_lm_nl", self.params["emb_size"], 2, self.params["num_landmarks"],
"sentence_embed", "lang_lm_mentioned"))
if self.use_aux_traj_on_map:
self.add_auxiliary(PathAuxiliary2D("aux_path", "map_b_r_select", "traj_gt_r_select"))
if self.use_aux_reg_map:
self.add_auxiliary(FeatureRegularizationAuxiliary2D("aux_regularize_features", None, "l1",
"map_s_w_select", "lm_pos_map_select"))
self.goal_good_criterion = GoalPredictionGoodCriterion(ok_distance=3.2)
self.goal_acc_meter = MovingAverageMeter(10)
self.print_auxiliary_info()
self.action_loss = ActionLoss()
self.env_id = None
self.prev_instruction = None
self.seq_step = 0
def __init__(self,
run_name,
ignore_lang=False,
class_loss=True,
ground_loss=True):
super(ModelTopDownPathGoalPredictor, self).__init__()
self.run_name = run_name
self.model_name = "top_down_path_pred_pretrain"
self.writer = SummaryWriter(log_dir="runs/" + run_name)
self.ignore_lang = ignore_lang
self.class_loss = class_loss
self.ground_loss = ground_loss
# The feature net extracts the 2D feature map from the input image.
# The label_pool down-sizes the ground-truth labels, which are input at the same size as the input image
# The output predicted labels are the size of the feature map
self.feature_net = ResNet13Light(32, down_pad=True)
self.label_pool = nn.MaxPool2d(8)
if self.ground_loss:
self.lang_filter = MapLangSemanticFilter(sentence_embedding_size,
32, 3)
self.aux_ground_linear = nn.Linear(3, 2)
enable_weight_saving(self.lang_filter, "ground_filter")
enable_weight_saving(self.aux_ground_linear, "ground_aux_linear")
if RESNET:
self.unet = ResNetConditional(sentence_embedding_size, 35, 2)
else:
unet_c_in = 35 if self.ground_loss else 32
unet_hc1 = 48 if self.ground_loss else 48
unet_hb1 = 24 if self.ground_loss else 24
self.unet = Unet5ContextualBneck(unet_c_in,
2,
sentence_embedding_size,
hc1=unet_hc1,
hb1=unet_hb1,
hc2=128,
split_embedding=splitemb)
if attention:
self.sentence_embedding = SentenceEmbeddingSelfAttention(
word_embedding_size,
lstm_size,
sentence_embedding_layers,
attention_heads=attention_heads)
else:
self.sentence_embedding = SentenceEmbeddingSimple(
word_embedding_size, sentence_embedding_size,
sentence_embedding_layers)
self.gather2d = Gather2D()
if self.class_loss:
self.aux_class_linear = nn.Linear(32, 64)
enable_weight_saving(self.aux_class_linear, "class_aux_linear")
print("Sentence Embedding #Params: ",
get_n_params(self.sentence_embedding))
print("U-Net #Params: ", get_n_params(self.unet))
print("Class auxiliary: ", self.class_loss)
print("Ground auxiliary: ", self.ground_loss)
# Enable saving of pre-trained weights
enable_weight_saving(self.feature_net, "feature_resnet_light")
enable_weight_saving(self.unet, "unet")
enable_weight_saving(self.sentence_embedding, "sentence_embedding")
if NLL:
#self.mask_loss = nn.BCELoss()
self.mask_loss = nn.NLLLoss2d()
elif BCE:
self.mask_loss = nn.BCEWithLogitsLoss()
elif CE:
self.spatialsoftmax = SpatialSoftmax2d()
self.mask_loss = CrossEntropy2d()
else:
self.mask_loss = nn.MSELoss()
self.aux_loss = nn.CrossEntropyLoss(reduce=True, size_average=True)
self.epoch_numbers = {"train": 0, "eval": 0}
self.iter = nn.Parameter(torch.zeros(1), requires_grad=False)
self.dropout = nn.Dropout(0.5)
self.dropout2d = nn.Dropout2d(0.5)
self.dropout3d = nn.Dropout3d(0.5)
self.viz_images = []
self.instructions = []
